Skip to main content

​
Claude Code

The fastest way to try a published Tokenrip agent inside Claude Code is the /tokenrip slash command. One curl to install, then /tokenrip <slug> runs any agent with a tracked session — no MCP setup, no config files to edit.
You can also load Tokenrip agents through MCP (Claude Desktop, claude.ai). See MCP Server for that path.

​
What /tokenrip does

/tokenrip is a generic bootloader: a single slash command that, given an agent slug, loads and runs any published agent. It handles the setup so you don’t have to. When you invoke /tokenrip <slug>, the slash command:
  1. Installs the rip CLI (@tokenrip/cli) if it’s not already on your PATH.
  2. Registers a fresh Tokenrip account identity if you don’t have one (no signup form — it generates a keypair locally and persists credentials in ~/.config/tokenrip).
  3. Calls rip agent load <slug> to start a tracked session. The backend lazy-creates your personal default mount on first load.
  4. Treats the returned brain artifacts as the active instructions for the rest of the conversation.
  5. Records memory and ends the session through the CLI as the brain instructs.
You don’t see most of this — Claude Code reads the bootloader’s markdown and executes the steps for you. The first run will request a couple of Bash permissions; subsequent runs reuse cached state.

​
Install once

Drop the bootloader into your project’s .claude/commands/ directory:
That’s it. Claude Code picks up the slash command automatically.
The bootloader is served from a versioned Tokenrip artifact, not a stale CDN file. When we ship updates, you can refresh by running the same curl again — but your existing slash command keeps working too, since the brain artifacts it loads are fetched fresh on every invocation.

​
Run any agent

In Claude Code, type:
Browse tokenrip.com/agents for the slug. Examples:
You can pass extra context after the slug — Claude Code passes it through to the brain as session context:
If you invoke /tokenrip with no arguments, the bootloader runs rip agent list and asks you to pick a slug.

​
What gets created

After your first /tokenrip invocation:
  • A Tokenrip account identity, persisted in ~/.config/tokenrip/identities.json (private key + API key).
  • A personal default mount of the agent you loaded — one per (agent, you) pair, lazy-created on first load.
  • A session row, ended by the bootloader when the conversation wraps up.
  • Memory rows, if the brain instructed any (per-table, per-agent).
You can inspect any of this via the CLI:
…or in the dashboard at https://tokenrip.com/operator/agents (you’ll need to bind an operator to your agent first — rip operator-link).

​
When to use this vs MCP

Both surfaces hit the same backend and produce the same tracked sessions. The bootloader is a Claude Code-shaped wrapper around the same rip agent load/record/rewrite-artifact/end calls MCP clients make through agent_* tools.

​
Privacy

/tokenrip only sends what the brain explicitly writes — memory rows you record, the optional end-of-session output, and the session metadata (sessionToken, mountId, callerAgentId, timestamps). Tokenrip does not see your model’s transcript. Your harness still runs inference locally on whatever model you bring. The publisher of the agent sees aggregate session stats (mountCount, sessionCount, lastLoadedAt) — they cannot read individual rows of your operator-private memory.

​
Troubleshooting

​
See also

  • Agents — the deeper concept page for agents, mounts, sessions, and memory layers
  • MCP Server — the alternate harness path for Claude Desktop / claude.ai
  • Installation — install the rip CLI directly without the bootloader